EEG-Based Epileptic Seizure Detection and Cross-Patient Generalization

Epileptic seizures occur when brain nerve cells exhibit abnormal electrical activity, sometimes causing loss of consciousness and convulsions. While epileptic seizures can be evaluated through EEG, they require long-term visual observation by highly trained specialists. Therefore, in emergency situations where specialists cannot respond, seizure detection may be delayed, leading to severe consequences. The greatest challenge in making automated detection practical is that EEG characteristics differ substantially between patients. Approaches that collect data and calibrate for each patient cannot be applied to a new patient immediately. Noting that the nature of inter-patient variability differs between seizure and non-seizure states, we extract patient-invariant features through domain adversarial training conditioned on the brain state. Furthermore, because a seizure evolves over time from its onset through the ictal period and into recovery, we also model how the extracted features change over time, making it possible to capture progressions that cannot be seen from individual time points alone. The goal of this approach is to complete training without using any data from the target patient. Because there is no need for a preliminary data collection period for each new patient, we believe it leads to a clinically practical system in which seizure monitoring can begin as soon as it is introduced. Technical approaches: - Extraction of patient-invariant features via state-conditional domain adversarial training - Modeling of temporal seizure progression with CNN and bidirectional LSTM - Seizure detection using features derived from stochastic generative models We are advancing verification using real clinical data through joint research with Okayama University Hospital.